This paper presents a real-time, pixel-wise method to generate grasp synthesis based on fully convolutional neural networks (FCN). Our proposed Attention Grasping Network (AGN) applies a novel attention mechanism to robotic grasp detection, which automatically learns to focus on salient features of the input image. The model with attention mechanisms can compensate for the loss of detailed information in standard FCN, which increases the sensitivity of the model and accuracy of prediction. In addition, in order to ensure a real-time grasp and save computing resources, the lightweight AGN model predicts the position and angle of the grasping point. Our method only takes 22ms to execute the grasp detection pipeline on a GPU-equipped computer and achieves 97.8% accuracy on Cornell Grasping Dataset.